I had dinner recently up in the Lehigh Valley with an old friend.

They’re a physician who also manages hotels. Probably a recession indicator when white-collar workers have multiple jobs, but that’s out of scope.

We were talking about software and this current wave of chatbots.

In my case, I predominantly use the web-interfaces and coding agents that live in my terminal.

In their case, they also use the dot com sites, but in addition, they have specific vendors that have an LLM integrated.

Despite the difference in domains, we have similar usage patterns.

Start with a good idea of the output we want. Use the clanker as a cognitive powertool. And validate the output against our own knowledge-base.

Effectively cutting a corner to reallocate that time elsewhere.

You can infer a lot about the priorities of the stakeholder by analyzing your user experience.

For example, let’s step back to the decay of Google search.

I remember the engine wars. Yahoo. MSN. AskJeeves.

In middle school, I was using GahooYoogle to have both sets of results showing side-by-side.

But by the time I was in high school, Google had won.

It was a better product than the competition. Within a single query, you could get the information you were looking for quickly. Without filtering on the user’s end.

Apparently that wasn’t good enough. Not enough “engagement” (read not enough clicks for advertisers). And so an internal battle was fought, resulting in the enshittification of search where ads and SEO slop polluted the first page results.

What does this tell us about the priorities?

Arguably this is a major contributing factor to what made the initial launch of ChatGPT feel like such a breath of fresh air.

So back to the chatbots.

Both my friend and I have noticed that when we prompt, the response oftentimes ends in a question. Ironically Gemini is the biggest culprit of this. Our expectation is that we use the tool to get the information we need, make changes, and move on to the next task. Which is at odds with the Silicon Valley obsession with the idea of “engagement”.

And so the developers and designers are incentivized by a paycheck or PIP to build a product that optimizes for that metric, to train a model with reward circuitry that gives it positive signal whenever a user replies.

Not an ad, but the Claude models have been pretty good at not doing this.

If I don’t want my DeWalt drill to ask me how driving screws into my deck felt, then why in Ishwar’s good name do I want the bot to ask me.